IB科学IA 数据处理 不确定度 误差棒 斜率
IB 科学的科学调查从 2025 年 5 月首评起换了准则，五项变四项，分数结构也跟着变了。旧大纲里的 Analysis 现在叫 Data analysis。名字换了，考的东西没换：数据够不够、处理对不对、不确定度算没算、结论能不能由数据支撑。IBO 官网的科目更新页写得很清楚，科学调查仍是 10 小时、占最终成绩 20%、报告上限 3000 词。这篇把数据分析这 6 分拆开讲，从表头怎么写一直讲到梯度不确定度。
The scientific investigation in IB sciences is worth 20% of the final grade, runs for 10 hours and caps the written report at 3,000 words, according to the IBO subject update pages. Since first assessment in May 2025 the five old criteria have been replaced by four, with data analysis carrying six of the 24 marks while conclusion and evaluation take the other half. What examiners actually reward in the data analysis section has barely changed, which is why old habits still decide most of the marks.
新纲四条准则里数据分析占 6 分
在 2025 年 5 月首评的版本里，24 分分给四块：研究设计、数据分析、结论、评价，每块 6 分。IBO 官网在生物科目更新页里点明，改版后有一半分数落在结论与评价上。旧大纲是五条准则，个人投入 2 分、探索 6 分、分析 6 分、评价 6 分、表达 4 分。名称和分值都动过，数据分析那 6 分的位置始终没动。很多学生以为改了准则就要换写法，实际上评分描述里对数据的要求几乎一字未改。真正变的是合作形式：小组成员可以共用同一套方法，自变量或因变量必须不同，数据必须是各自的。
Four criteria now carry 24 marks: research design, data analysis, conclusion and evaluation, six marks each. The IBO subject updates state plainly that half of the available marks sit in the conclusion and evaluation, a shift from the older model in which analysis was one of five roughly equal blocks. Group work is permitted where students share a method but differ in variables, and the data each student submits has to be their own.
表头要写单位和仪器的不确定度
数据处理的第一处失分几乎都出在表头。一列原始数据的标题只写质量、长度、时间，阅卷时看不出你用的仪器有多精确。规范写法把量、符号、单位和不确定度一起写进去：长度 l / cm ± 0.1。模拟刻度取最小刻度的一半，量程 30 厘米的尺子最小刻度 1 毫米，读数不确定度记 0.5 毫米。数字仪表直接取末位一格，电子天平显示 12.34 克，不确定度就是 0.01 克。常见的错误是把分辨率当成不确定度，或者干脆不写。表格里还要有重复次数的说明，三次读数取平均时，平均值的随机不确定度可按极差的一半估算。
Table headings carry real marks, and the ones that lose them are almost always missing either a unit or an uncertainty. Write the quantity, its symbol, the unit and the instrument uncertainty together, as in length l / cm plus or minus 0.1; an analogue scale is normally read to half its smallest division, while a digital display is quoted to its last digit. Repeats belong in the table too, because a mean of three readings carries a smaller random uncertainty than any single reading, and saying so in the heading saves the reader from guessing.
有效数字跟着最不精确的那个量
有效数字不是越多越好，处理后的结果不能比输入里最不精确的量更精细。用最小刻度 1 毫米的尺子量出 12.3 厘米，再乘上给定到四位有效数字的密度，最终体积只能保留三位。很多学生以为计算器显示几位就写几位，答案栏里出现 8.249371 这样的数，表达分直接丢掉。中间步骤可以多留一位，避免累积舍入误差。最终结果的有效数字要和不确定度对齐：不确定度写在末位，比如 9.81 ± 0.05 米每二次方秒，写成 9.8123 ± 0.05 就自相矛盾。
Significant figures are decided by the least precise measurement in the chain, not by what the calculator shows. Carrying one extra digit through intermediate steps is good practice, yet the final value has to agree with its own uncertainty, so 9.81 with an uncertainty of 0.05 is quoted to two decimal places and never to four. Examiners see long strings of digits in the answer line every session, and it reads as if the uncertainty was added as an afterthought.
不确定度传递：加减看绝对值，乘除看百分比
把两个读数相减，不确定度相加而不是相减。量程 100 厘米的尺子测两段长度，各自 ± 0.1 厘米，差值的不确定度就是 ± 0.2 厘米。乘除时把每一项的相对不确定度换算成百分比再相加，最后乘回结果。举一个能算完的例子：质量 2.00 ± 0.01 克，体积 5.0 ± 0.1 立方厘米，密度的相对不确定度是 0.5% 加 2%，约 2.5%。幂次要把指数提出来当乘数，平方项的相对不确定度翻倍。常见的错误是只报原始读数的不确定度，处理后的量就不再交代。表格最后一列写出最终值加减不确定度，核对成本立刻降下来。
Adding or subtracting readings means adding their absolute uncertainties, while multiplying or dividing means adding percentage uncertainties; a power pulls its exponent out as a multiplier, so squaring a quantity doubles its relative uncertainty. Work one example through in full, for instance a mass of 2.00 with an uncertainty of 0.01 and a volume of 5.0 with an uncertainty of 0.1, which gives a density uncertainty near 2.5% rather than the 0.5% carried by the mass alone. Leaving the processed quantity without any uncertainty is the most common gap in this section.
把曲线改成直线再求梯度
研究问题里的关系大多不是线性的，直接作图得到一条弯线，梯度和截距都无从谈起。单摆是最好的例子：周期 T 与摆长 L 满足 T = 2π√(L/g)，以 T 为纵轴画出来是曲线。把纵轴换成 T 的平方，图变成过原点的直线，梯度等于 4π²/g，用梯度反解 g 只需一步。改写坐标轴是被允许的，前提是写清为什么这样改，并在图上标出新的量。很多学生以为换轴就是凑一条直线，其实线性化本身就是把物理关系写进坐标轴的一步推导。另一类常见处理是取倒数：速率与时间的关系改成速率对浓度作图，斜率直接对应速率常数。
Most relationships in an investigation are not linear, and a curved set of points gives neither a useful gradient nor a meaningful intercept. Linearising an axis is a legitimate move as long as the transformation is stated and justified, as with a pendulum where plotting the square of the period against length turns the curve into a straight line through the origin whose gradient is four pi squared over g. Explain the substitution in one sentence, label the new axis with its unit, and the gradient becomes a route back to the constant you wanted.
最佳拟合线不是把点连起来
评分描述里对作图的要求很具体：点要用清晰的小叉或实心点标出，不画折线把相邻点连起来。最佳拟合线是一条平滑的直线或曲线，让点均匀分布在两侧，个别偏离明显的点不参与拟合但要标出来。用软件拟合时把方程和决定系数一起写在图上，决定系数只说明线性程度，替代不了对物理意义的判断。常见的错误是在明显弯曲的趋势上硬拉一条直线，或者把线强行画过原点。截距有物理意义时也要读出来并给不确定度，系统误差往往就藏在一个不为零的截距里。图上的网格线和字号要顾及投影效果，印在纸上看得清、投到屏幕上糊成一片的图不加分。
A line of best fit is not a join-the-dots exercise; the points are plotted clearly, the trend is drawn as a single smooth line, and a point that is visibly off is flagged rather than quietly absorbed. Software fits should come with the equation and the coefficient of determination on the graph, though a high value only says the line fits, not that the physics behind it is right. Read the intercept as well, because a non-zero intercept where theory predicts zero is usually the first sign of a systematic error.
斜率和截距的单位决定结果对不对
梯度不是一个纯数字，它带着纵轴单位除以横轴单位。纵轴伏特、横轴安培，梯度的单位就是欧姆，不是伏特。数据处理里最容易忽略的一步，是把坐标轴单位统一到国际单位制再算。毫克换成千克差 10 的负 6 次方，毫秒换成秒差 10 的负 3 次方，梯度会因此差好几个数量级。求梯度不确定度的通行做法，是从误差棒两端各画一条最陡和最平的可行拟合线，两条线梯度之差的一半就是不确定度。手算也行，取拟合线上相距较远的两点算斜率，两点越远相对误差越小。把梯度写成数值加减不确定度再加单位，这一步才算交代完整。
A gradient is never a bare number, it carries the vertical unit divided by the horizontal unit, which is why a volt-against-amp graph yields ohms. Converting everything to SI units before fitting prevents the factor-of-a-thousand slips that come from milligrams and milliseconds, and those slips are hard to spot once the graph itself looks convincing. The standard way to quote a gradient uncertainty is to draw the steepest and shallowest acceptable lines through the error bars and take half the difference between their gradients.
误差棒重叠不代表差别不显著
误差棒画出来以后，很多学生以为只要两组数据的误差棒有重叠，就可以写下没有显著差异。这个判断在统计上并不成立，重叠的幅度和样本量都会影响结论。更稳的做法是看两个均值之差与合成不确定度的比较，差值大于合成不确定度两倍时，通常可以认为差别超出了随机波动。写报告时把判断依据写出来，比只给一句结论更有说服力。误差棒本身也要说明代表什么：标准差、标准误还是仪器的极限误差，三者含义完全不同，混用会让整段分析失去意义。图注里一句话说清楚就够了，不必在正文重复解释。
Overlapping error bars are routinely read as proof that two results do not differ, and that inference is unsafe because overlap depends on the size of the bars and on how many readings sit behind them. A more defensible line is to compare the difference between the means with the combined uncertainty, and to spell out the rule you applied rather than asserting a verdict. State what the bars represent as well, since a standard deviation, a standard error and an instrument limit are three different claims about the same data.
异常点什么时候可以剔除
剔除数据是允许的，前提是理由写在报告里，而不是图上悄悄少一个点。可以接受的理由包括读数记录错误、仪器在某一刻失灵、操作步骤明显偏离。不能接受的理由是它让结果不好看。规范做法是把异常点保留在表里并用脚注说明，同时给出包含与不包含它的两套拟合结果，让读者自己判断影响有多大。常见的错误是删掉点以后不提，或者把偏离点当成新发现去解释。三个重复读数里有一个明显偏离，先查记录和装置，再决定去留。阅卷时看到图上少一个点而正文没有交代，这一段的分基本就停了。
Removing an anomalous point is legitimate only when the reason is stated, such as a misread instrument, a known fault during one run, or a step that clearly departed from the method. Keep the point in the table, footnote it, and show the fit with and without it so a reader can judge how much the conclusion depends on that decision. Deleting a point and saying nothing is the version that costs marks.
百分比差异和百分比误差不是一回事
把实验结果和文献值比较时，两种算法常被混用。百分比误差以公认值为分母，衡量实验值偏离真值多少；百分比差异以两者的平均值为分母，衡量两个测量彼此差多少。分母不同，数字会差出几个百分点，结论的口气也跟着变。报告里写清用的是哪一种，并给出文献值的确切来源和年份。文献值不是随手从网上抄的数字，教材后的数据表、考试局的数据手册、原始论文都能引，但必须写出处。比较之后要解释差异的方向：系统偏大还是随机分散，系统偏大通常指向方法，随机分散指向读数精度。这句解释往往比百分比本身更重要。
Percentage error and percentage difference are not interchangeable, because the first divides by the accepted value while the second divides by the mean of the two measurements. Choose one, name it, and cite where the accepted value came from along with its year, since a number lifted from a website without a source carries no weight in moderation. Then explain the direction of the discrepancy, because a consistent bias points at the method while scatter points at reading precision.
计数数据用卡方，比较均值用 t 检验
生物和心理的 IA 常需要一句统计判断，用错检验比不做检验扣得更多。分类计数适合卡方检验：把观察频数与期望频数逐格算差，平方后除以期望值再求和。自由度等于类别数减一，查表时取 0.05 的显著性水平，算出的值超过临界值就认为偏离显著。连续数据比较两组均值更适合 t 检验，样本量小的时候它对分布形状更敏感，报告里要写清样本量和假设。常见的错误是把百分比直接拿去做卡方，或者把三次读数当成三个独立样本。检验只是工具，动手之前先想清楚研究问题问的是什么，再选方法。
Categorical counts belong to a chi-squared test, where each observed frequency is compared with the expected one, squared, divided by the expected value and summed, with degrees of freedom equal to the number of categories minus one. Comparing two means from continuous data calls for a t-test instead, and with small samples the assumptions about distribution need to be stated rather than silently assumed. Running the wrong test, or treating three repeats as three independent samples, is worse than presenting the data without any test at all.
从 4 分到 6 分差在哪里
同一条准则，4 分和 6 分的差别不在数据量，而在处理与解释是否完整。4 分档常见的样子是：数据够用，处理步骤大致正确，不确定度提到了一部分，解释与结论只搭上一半。6 分档要求处理全程正确，不确定度在每一步都跟到，解释与研究问题直接对应，并且能指出数据本身的局限。很多学生以为多做几组数据就能上去，缺的通常是解释那一层：图有了，没读出趋势代表什么；数字有了，没说明它为研究问题提供了什么证据。把每一张图的读数写成一句结论，是最省力的提分动作。
The gap between a middle mark and full marks is rarely about how much data was collected, it is about whether processing and interpretation are both complete. A report in the middle band usually has enough data and roughly correct processing, with uncertainties mentioned in some places and an interpretation that answers half the research question. Full marks need every step correct, uncertainties carried through the whole chain, and an explicit statement about what the data can and cannot support.
3000 词里数据分析和图表占多少
报告上限 3000 词，图表和表格不计入词数，这个规则值得用足。把预算分开会更从容：研究设计与背景约 600 词，方法与安全约 300 词，数据处理与图表说明约 900 词，结论与评价约 900 词，剩下的留给引用与说明性文字。数据分析这一块往往写得最薄，因为多数人把功夫花在实验过程上。图表数量控制在四到六张，每张配一句说明它证明了什么，比十张没有解读的图得分高。写完之后通读一遍，把所有形容词删掉，看剩下的句子是否仍然说清同一件事。删不掉的才是内容，删掉后句子塌了的地方，多半是原句本来就没写清。
With a hard cap of 3,000 words and tables excluded from the count, a working split is roughly 600 words on research design and background, 300 on method and safety, 900 on data processing with figure commentary, and 900 on conclusion and evaluation. Data analysis is usually the thinnest section in real reports because the effort goes into the practical work rather than the write-up, and four to six well-commented figures score better than ten unexplained ones. Read the draft once with every adjective removed to see whether the substance survives.
TI-Nspire CX II CAS 图形计算器 AP/IB/A-Level¥1320📱 去京东购买京东精选
需要一对一辅导？Need One-on-One Tutoring?
A-Level / GCSE / IB / KS2-KS3 全科辅导
16621398022 同微信
Visit aleveler.com
或截图后用「扫一扫 → 相册」选择本图