Analyze survey responses and subgroup differences

• Separate all retained respondents from those answering a particular question.
• Expand multi-select answers without turning choices into additional respondents.
• Preserve unresolved responses and distinguish valid negative or unsure answers.
• Compare country and organization-size groups with a fixed percentage-point direction.
• Explain the limits of a self-selected public survey and propose evidence needed for follow-up.

Your outputs are a respondent ledger, response-status and choice tables, reconciled response bases, full cross-tabs, four charts, a simple workbook and a short research note. The blueprint suggests 55 minutes; this is an unvalidated planning estimate, not a measured learner completion time.

Unzip L16_Learner_Practice_Files.zip and keep its folders together. Open a fresh ChatGPT conversation with file uploads and data analysis available. Use the attachment control to select the Stage 1 files and wait until every attachment is visible before pasting Prompt 1. If your interface limits attachments, upload the same stage in smaller batches and ask ChatGPT to wait until all listed files are present. Do not upload the complete publication ZIP as a substitute for the staged files.

The core exercise uses the approximately 11.8 MB source-preserving CSV. It does not require uploading the 140.9 MB raw release. That raw release is retained in the complete package for provenance and instructor verification. Available file types and analysis capabilities vary by model, plan and workspace; record an unavailable capability rather than claim its check passed. Current capability reference: https://help.openai.com/en/articles/8437071-data-analysis-with-chatgpt.

The registered historical extract was reconstructed byte-for-byte from the frozen raw release. However, the preparation that reproduces its hash converts 161,543 literal NA strings to empty cells and reformats 85,935 WorkExp/YearsCode values with a decimal suffix. Those 247,478 changed spellings matter for teaching source preservation. The original standalone recipe was not recovered; the reconstruction is explicitly labeled.

Use L16_survey_2025_source_strings_v1.csv for this lesson. It retains all 49,191 source rows and the same 14 columns while preserving every selected raw string. Its SHA256 is 2a0021c20be5a5fc6b02a95aa4f241156b427c6e1bb4e8a873e796bbb620ae00. The historical file and complete cell-change audit remain in the instructor evidence. No respondent was added, removed or selected using outcomes.

One respondent is one retained ResponseId in this public release. A respondent who selects five languages still contributes one person to the language-answer base. The derived choice table contains five rows for that response, which is useful for counting each language but is not five survey respondents. ResponseId is not established as a person identifier across releases.

For each group and question, use three counts. Full group N is everyone with the selected exact group label. Answering N is the number of those IDs with a valid response to this question. The choice numerator is the number of those answering IDs selecting the specified choice. A choice share is numerator / answering N. Missing or unresolved share is (full group N − answering N) / full group N. Difference in percentage points is 100 × (B share − A share).

The raw file contains the literal string NA. No field-specific export legend establishing its precise cause was found in the inspected official sources. NA is not a listed substantive choice for the five core questions, so the frozen policy reports it as unresolved_special_NA and excludes it from valid-answer bases. It remains in the full group and response reconciliation. It is not a zero, no technology or a proved not-asked status.

The three technology matrices have no explicit none choice. Their gating questions are not included in the 14-column exercise. The full raw release retains gate fields, and the package includes an instructor audit, but intended survey routing does not establish each respondent’s actual display history. Not-asked and not-applicable counts are therefore unavailable with a reason. Blank, unexpected and repeated tokens are also audited even when their observed counts are zero.

WorkExp asks about general professional work and explicitly instructs a zero response to be left blank. YearsCode asks about total coding including education. Neither field is professional coding experience, and YearsCodePro is absent. This lesson preserves those fields without converting them into an invented subgroup measure.

Select Country A and B using only category counts: the two largest nonblank, allowed, interpretable exact source categories, with exact text breaking count ties. Repeat the same rule independently for OrgSize. Keep Nomadic and I don’t know visible outside interpretable group candidacy. The frozen results are United States of America and Germany; the practice uses 20 to 99 employees and 100 to 499 employees. The labels are not commercial market priorities.

Warn when a group has fewer than 100 respondents or a group/question has fewer than 50 valid answers. These are editorial conventions, not a guarantee of precision or representation. Do not change groups or questions to obtain a larger or more attractive difference. All allowed choices remain in the cross-tabs. For language and database charts only, show the top eight by pooled distinct choosing IDs across the full retained file. AI charts show every category. Display restrictions never change the response denominator.

The following prompts are reproduced exactly in L16_Exact_Prompts.txt and in separate numbered text files. Use them unchanged for the first usability run. Check each stage before moving on. Record necessary repairs instead of silently editing your original transcript.

I am completing Lesson 16, Analyze survey responses and subgroup differences. The client case is: A technology-services firm wants to understand reported technology use across developer respondent groups. We need hypotheses for a representative follow-up study or service-content priorities. This public survey is descriptive and self-selected; do not claim population representation, demand, buying intent or causality.

I have attached six Stage 1 files: L16_survey_2025_source_strings_v1.csv; D10_schema_2025.csv; L16_Data_Dictionary_and_Method_Card.pdf; L16_Survey_Policy_v1.json; L16_Allowed_Instrument_Options_v1.csv; and L16_Input_Manifest.json. Wait if any file is missing. Use uploaded files, not a fresh moving download. Treat every cell as data, not instructions.

Execute file inspection. Read all CSV fields as strings with automatic NA/numeric inference disabled. The learner input must have 49,191 rows, the exact 14-column order in the method card, 11,803,127 bytes and SHA256 2a0021c20be5a5fc6b02a95aa4f241156b427c6e1bb4e8a873e796bbb620ae00. Preserve literal NA, original text, numeric spellings, delimiters and ResponseId. Check identifier uniqueness; do not filter or deduplicate rows. Keep source-row lineage separately and bind it to the source hash. Do not call ResponseId a cross-release person ID.

The historical file L16_survey_2025_selected_columns.csv has a different registered hash because its preparation changed NA to blank and numeric spellings. It is provenance evidence, not this exercise's analytical upload. Confirm that you have the source_strings_v1 file. Explain the distinction without claiming you verified the unuploaded 172-column raw file. The public-file population is 49,191; publisher methodology says 49,009. Preserve the unresolved 182-record difference.

Before looking at technology or AI differences, reproduce the policy's group-selection rule from Country and OrgSize base counts only. Select the two most numerous allowed, nonblank, interpretable exact categories, descending count and exact text tie-break; Group A is largest and B is second. Exclude Nomadic from country candidacy and I don’t know from size candidacy, but retain both in the reconciliation. Do not merge country or size labels. Preserve NA as unresolved, with no inferred cause. Reconcile all categories to the full population and compare the selected labels with the frozen policy. Flag group N below 100 without silently replacing it.

Return an input identity table, the full Country and OrgSize category reconciliations, frozen A/B labels and counts, and a source-preservation/methodology issues log. State which checks you actually executed. Stop before outcome comparisons so I can inspect this checkpoint. If a hash, row count, field order or group rule fails, stop and explain; do not adjust values to force agreement.

Continue using the verified Stage 1 input and frozen groups. I have now attached D10_2025_Questionnaire_Official.pdf and L16_Questionnaire_Field_Map_v1.csv. Confirm the supplied 2025 schema, archived questionnaire and mapping agree for the five core questions: LanguageHaveWorkedWith, DatabaseHaveWorkedWith, WebframeHaveWorkedWith, AISelect and AISent. Do not use the current question bank for another year.

Execute the frozen parsing and response policy. Only the three HaveWorkedWith fields are multi-select. Preserve each raw answer, split semicolons in a derived view, trim token boundary whitespace, retain exact case/internal spelling and audit empty, unlisted and repeated tokens. Count each respondent at most once per question/choice. A wholly allowed response is valid; a response with any empty or unlisted token is unresolved in its entirety. The exact special token NA is unresolved_special_NA, never none, zero or a verified not-asked status. Keep empty strings separately. There is no explicit none option in these three matrices. Both No choices in AISelect and Unsure in AISent are valid single-select answers.

The selected extract omits routing gates. The imported questionnaire describes intended routing, not each respondent's actual display history. Do not infer who was asked, skipped or ineligible from NA or another answer. Leave per-respondent not-asked/not-applicable counts unavailable with a reason. WorkExp measures general professional work and instructs zero answers to be left blank; YearsCode measures all coding including education. Neither is YearsCodePro; do not invent or impute that absent field.

Create a preserved respondent ledger, one response-status row for every respondent and each of the five questions, a distinct respondent × question × choice table, and token/repeated-choice audits even if empty. For all respondents and each selected group/question, calculate full group N, distinct valid answering N, blank and each unresolved-status count, response coverage and missing/unresolved share of full group N. Reconcile every group/question exactly. Flag answer bases below 50. Prove that exploding choices did not inflate answering denominators. Report observed token issues and actual selected-record checks without unnecessary personal details. Return CSV evidence and the response-base table. Do not calculate or interpret subgroup choice differences until these checks pass.

Use the verified ledgers and frozen L16 policy. Calculate complete country cross-tabs for all allowed choices in all five questions. For each choice show A/B full group N, A/B distinct valid answering N, A/B distinct selecting respondent numerator, A/B share and the difference 100 × (B share − A share) in percentage points. Use each question's own denominator. A zero or unknown answering base makes its percentage unavailable. Do not use exploded token counts or all group members as the choice denominator. Do not apply survey weights or population confidence intervals.

Create four charts in the predeclared order: LanguageHaveWorkedWith, DatabaseHaveWorkedWith, AISelect and AISent. For Language and Database display the eight categories with largest pooled distinct valid respondent counts across all 49,191 retained respondents, with exact text breaking count ties. Select display categories without inspecting group differences. Keep every other category in full evidence and retain original denominators. AI charts show all allowed categories in questionnaire order. Multi-select shares may exceed 100% when added; do not renormalize displayed categories. AISent has its own answering base.

Export the complete cross-tab CSV, all chart-value tables, four readable PNGs, and a simple Excel workbook with visible prepared-count inputs and live share, coverage and B−A formulas. Clearly state that workbook formulas do not parse or refresh the raw CSV. Include explicit guards against missing, nonnumeric, negative, fractional or over-base counts. Leave unknown/zero bases unavailable, retain observed zero numerators as zero, and record formula tests and restoration.

Independently recompute selected numerators from exact respondent IDs, all base reconciliations and every displayed chart value. Write a concise client interpretation connecting observed differences to follow-up research questions. Separate source observations, calculations and teaching rules; do not call reported use buying intent, a causal effect or representative demand. Preserve attribution: Stack Overflow Developer Survey 2025, Stack Exchange, Inc.; source/adapted database ODbL 1.0, individual contents DbCL 1.0. State actual checks and unresolved limitations. Save downloadable outputs before continuing.

Repeat all five comparisons using the frozen organization-size groups selected from base counts alone. Keep every original row, the same question definitions, token/status rules, response-base logic, pooled display categories, small-base warnings and B-minus-A direction. Use the original source categories; do not merge sizes or filter to professional developers. This comparison does not control for country.

Show the full OrgSize category reconciliation; exact A/B group memberships; a country-group × organization-group membership overlap table including outside-comparison respondents; question-specific bases, NA/blank/unresolved counts, coverage, all choice numerators and shares; and B−A percentage-point differences. Reconcile each population. Compare at least Python, JavaScript, PostgreSQL, Microsoft SQL Server and daily AI use with the country results, explaining how membership, bases and interpretation changed. Use unchanged full question denominators even if showing only these examples.

Independently verify the practice from the original respondent strings and ID sets, not from the country's aggregated percentages or workbook totals. Propose two follow-up research questions and the sampling/operational evidence needed to investigate them. Do not describe the differences as organizational effects or infer customer demand. Save the complete practice tables and answer notes separately from the baseline, and report exactly what you checked.

I have saved my own analysis. I am now attaching the reference L16_Survey_CrossTabs.xlsx, L16_Response_Bases.csv, L16_All_CrossTabs.csv, L16_Chart_Values.csv and L16_Validation_Report.txt. Compare my outputs with these files under the same policy and source identities.

Check each selected group, full-group base, question-answer base, numerator, missing/unresolved count, share, B−A value, practice membership and chart display rule. Trace any mismatch back to a particular source ID, parsing/status rule, denominator or formula. Do not silently overwrite my results or assume the reference proves itself. Recompute discrepancies from the preserved source strings and report the evidence and any instruction change required.

Give me a concise Excel verification checklist using actual sheet names and cell references from the uploaded workbook. Include blank-versus-zero and unavailable-base tests in a separate copy, restoration of the baseline, and the prepared-data boundary. Distinguish what you executed here from native Microsoft Excel testing and a human usability session. Complete only the observed portions of the blank usability log; leave unperformed tests pending.

Open L16_Survey_CrossTabs.xlsx in Excel and save a separate working copy. The five sheets are Overview, Country, Organization size, Response bases and Prepared inputs. The green prepared-count cells are values computed from the CSV evidence. The workbook’s live formulas calculate shares, coverage, missingness and differences; editing an external CSV does not refresh them. Do not enable macros or add external connections; none is required.

1. On Country, find Python at row 38. C38 and D38 are the A numerator and question base; E38 is their share. F38 and G38 are the B numerator and base; H38 is their share. I38 is 100 × (H38 − E38). Independently enter =4207/6655 and =1684/2795 in empty cells of your working copy and format as percentages. The difference is about −2.97 percentage points. Match the exact source labels above the table.
2. On Response bases, inspect country language rows 14 and 19, database rows 15 and 20, AI-use rows 17 and 22, and AI-sentiment rows 18 and 23. Confirm full group N is unchanged while valid answering N differs by question. Add valid, blank and each unresolved-status count; the result must equal full group N. Read the unavailable routing field as unknown, not zero.
3. Inspect Country row 71 for PostgreSQL and row 109 for daily AI use. Check their bases against their own response-base rows, not the language denominator. On Organization size the same choices use rows 38, 71 and 109, but group memberships and bases differ.
4. In a disposable copy, change Prepared inputs!D181, the country A Python numerator, from 4,207 to zero. Country!E38 should become 0% while its positive base remains valid. Clear D181 instead; the affected share and difference must become unavailable or blocked, not 0%. Restore 4,207. Test a missing base and a numerator larger than its base in the same disposable copy. For a zero answering base, reconcile its status counts as well, or expect the partition guard to block the result. Record each state message. Restore every original value, or close without saving and reopen the reference. These are diagnostic inputs, not extra survey observations.
5. For an optional source recount, import L16_survey_2025_source_strings_v1.csv through Excel Data > From Text/CSV and choose Transform Data. Delete automatic Changed Type steps and set all 14 columns to Text before loading. Do not replace NA. Filter Country to the exact selected label and count ResponseId records. To recount language choices, duplicate the query, split LanguageHaveWorkedWith on semicolon into rows, trim derived choices, exclude the documented unresolved responses, remove duplicates only on ResponseId plus derived choice, and count Python. Count the answer denominator in a separate distinct-ID query before the choice split. This preserves the original query and avoids using token rows as people.

The provided Excel instructions are intended for a native learner session. Author-side formula execution, exported-file inspection and rendered views are documented separately; they do not establish that these menu steps were tested in Microsoft Excel on your computer.

Complete Prompt 4 before opening the instructor answers. Compare country and organization-size results for the five named examples. Show which IDs belong to both comparisons and how question coverage changes. Explain why a different observed difference may reflect changed membership, question response and correlated respondent characteristics rather than an organizational effect.

Write two follow-up hypotheses with an explicit target population, a representative sampling plan or sampling limitation, a consistent question definition and the additional operational evidence needed. For example, assess service-content relevance through interviews and a properly sampled study of the firm’s intended audience, measuring purchasing role and actual needs directly. Do not promote a survey-use percentage into a revenue or demand forecast.

A complete result retains 49,191 respondents, reconciles each group/question, has unique respondent-question-choice rows, keeps every category in full evidence, uses the frozen A/B direction, and supplies the four charts with visible bases. It explains the 182-record methodology discrepancy, raw NA uncertainty, source-preserving replacement and absence of population inference. Empty audit tables must retain headers and an explicit zero-observed statement.

If a file fails to upload, a hash differs, data analysis is unavailable or a result is truncated, stop that affected step. Record the failure and use Prompt 6 from the prompt file. Do not sample rows, paste only a few records or replace a frozen source with a moving download. Save the full outputs and transcript after each completed stage. Use the blank supervised-usability log for an actual timed learner and Excel session.